xnn.gnn.models.gemnet

GemNet (Gasteiger, Becker, Guennemann, NeurIPS 2021, arXiv:2106.08903).

Geometric message passing over directed edge embeddings with one-hop (triplet) and two-hop (quadruplet) interactions, built on the xnn bases shared with DimeNet; one class covers GemNet-Q, -T, -dQ and -dT.

Functions

he_orthogonal_(weight)	He-scaled orthogonal initialization with exact standardization (paper section 6).
reverse_edges(edge_index, edge_vec)	The index of the reverse a -> c of every directed edge c -> a.
triplets(edge_index, num_nodes)	The triplets c -> a <- b of one-hop message passing (b != c).

Classes

GemNet([n_features, n_edge_features, ...])	GemNet: geometric message passing (Gasteiger, Becker, Guennemann, NeurIPS 2021).
ScaleFactor()	A constant factor that keeps the activation variance of a layer (paper section 6).